DE102021200938A1 - PROCEDURE FOR TESTING A FLUID MACHINE COMPONENT - Google Patents

Abstract

The present invention relates to a method for testing a component, in particular a component of a turbomachine (1), with the steps: - creating (S2) at least one X-ray or CT image of the component (10) by means of an image acquisition device (20); Providing (S21) metadata about the component (10), which include in particular a component type, a term of the component (10), a number of remaining life cycles and/or a repair history;- Classifying the component (10) using the image acquisition device ( 20) created recording and the metadata provided into a "servicable" or a "non-servicable" category using a machine learning system (30).

Description

technical field

The present invention relates to a method for testing a component of a turbomachine and a method for training a machine learning system.

State of the art

An axial flow machine is functionally divided into compressor, combustion chamber and turbine, whereby in the case of an aircraft engine, intake air is compressed in the compressor and burned in the downstream combustion chamber with added kerosene. The resulting hot gas, a mixture of combustion gas and air, flows through the downstream turbine and is expanded in the process. The turbine and the compressor are usually constructed in multiple stages, with each stage comprising a stator and a rotor. The stators and rotors are each made up of a plurality of circumferentially consecutive blades around which the compressor or hot gas flows, depending on the application.

In the course of use, an aircraft engine or a power plant turbine can suffer damage, such as damaged areas, structural changes or signs of corrosion on its components. The components arranged in the gas duct can be particularly endangered and relevant to safety, for example blades, but also gas duct plates etc. Even small damaged areas can be problematic insofar as they can represent an initial point for crack propagation. Therefore, the detection of such damaged areas can be of particular importance.

It is known in the prior art that during the so-called assessment of components such as blades, for example, images, in particular X-rays or CT images, are created from these. As a result, cracks or geometric deviations in components operated in the turbomachine or pores or insufficient welded connections in repaired components can be detected non-destructively.

There are already software solutions that mark probable flaws based on predetermined criteria and thus support the diagnostician in his work. However, the known software solutions are not yet sufficiently reliable, so that a manual check is necessary in any case.

Presentation of the invention

The present invention is based on the technical problem of specifying an advantageous method for testing a component, in particular a component of a turbomachine.

According to the invention, this is achieved with the method according to claim 1 . In this case, at least one X-ray or CT image is first created of the component by means of an image acquisition device. In a further step, metadata about the component is acquired or provided, which includes, for example, a component type/part number, a term of the component, a calculated remaining service life, a prescribed remaining service life, an operator of the turbomachine, a wall thickness or a repair history of the component. On the basis of the at least one first image created by the image acquisition device and the acquired metadata, the examined component is classified by means of a machine learning system, e.g. using a (for this purpose) trained self-learning algorithm that runs on a computer/computer, either under a "Servicable ' or a 'Non-Servicable' category. In this context, “servicable” means that the component is still serviceable, i.e. could be installed again in a turbomachine while complying with the approval conditions, while a component classified as “non-serviceable” is no longer serviceable and either requires repair or is scrap . In other words, the invention provides a method for testing a component using a machine learning system that has at least one computer or a similar computing device with a trained self-learning algorithm stored thereon. The machine learning system receives at least one X-ray or CT image of the component from an image acquisition device and at least one set of metadata about the component, e.g. from a database, and then evaluates them using the trained self-learning algorithm.

One advantage of this is that the component can be examined for defects, in particular cracks and/or pores, quickly, automatically and in a technically simple manner.

Advantageous configurations of the invention are the subject matter of the dependent claims.

According to one aspect of the invention, the machine learning system can be further trained to assign the parts classified as "non-serviceable" to either a "repairable" or a "non-repairable" category. This enables leaner workflows, since scrap parts (scrap), in the case that the machine learning system can make a clear classification can be sorted out quickly.

According to a preferred embodiment, the machine learning system can also be trained to assign a probability of success of the repair to the components classified as “repairable”. In particular, there can be feedback on the repairs actually carried out on components of the same type, i.e. the input as to whether a repair was successful is fed back into the meta database of the machine learning system and with the test data, i.e. the recording from the image acquisition device and the other metadata connected. With a sufficiently large pool of data, a link can be established between the type, number and size of the defects found and the statistical probability of successful repair for a specific component type or for all components. In a further preferred embodiment, if a tested component falls below a certain probability of repair, for example less than 60%, preferably less than 40%, the machine learning system can evaluate it as scrap. This helps to avoid uneconomical repair attempts and to use resources in the MRO more efficiently.

According to a further aspect of the invention, the machine learning system can comprise a neural network, a deep neural network (DNN), a convolutional neural network (CNN) and/or a support vector machine.

According to a further aspect of the invention, the machine learning system can be designed to identify and/or localize defects, in particular cracks or pores, contamination, connection errors in the at least one recording. The identified flaws, in particular a type, number, position and/or size of the identified flaws, can be taken into account in the classification of the component. For example, particularly critical positions of cracks can be identified, where component failure is particularly likely. This corresponds more to a symbolic approach of machine learning, i.e. as a "jump start" the solution is based on the procedure of a human appraiser/tester, who in a first step identifies and evaluates defects in the recording by hand.

In a further embodiment of the invention, the machine learning system can be designed to place a filter over the recording obtained from the image acquisition device as a classification step. This is another step towards a symbolic approach to get faster to a converging solution. For human examiners/assessors, some errors become clearer, e.g. when changing the gray level gradients. The machine learning system usually first looks at the entire, unfiltered data set of the recording coming from the image acquisition device. When training the machine learning system, if it receives the manual marking of the defect in the recording together with the filter setting chosen by the examiner/appraiser at which the defect was recognized, the machine learning system can also filter the data with this filter to simplify pattern recognition and possibly converge faster to a solution.

According to a further aspect of the invention, the machine learning system can be designed to autonomously control the image acquisition device after the evaluation of the at least one recording in order to create at least one further recording of the component with a varied recording parameter, in particular a varied recording angle, if based on the at least an initial admission a classification criterion cannot be met.

The object is also achieved by the method according to the invention for training a machine learning system that is designed and provided for testing a component, in particular a component of a turbomachine, according to claim 8. In a first step, a machine learning system is provided, which in particular is a neural network includes.

A recording of the component and metadata about the component are entered into the machine learning system, the metadata including in particular a component type, a term of the component, a number of remaining life cycles and/or a repair history.

In a further step, the machine learning system classifies the component based on the entered data into a "servicable" or a "non-servicable" category and then outputs the determined category, e.g. on a screen.

Appropriate information about the category of the component is then fed into the machine learning system to train it. This means that the machine learning system first checks the component on the basis of its current training or knowledge level and is then trained further with a selected "sample solution". If this sample solution or applicable information matches the evaluation of the machine learning system, you can continue with the next component. Are the test results correct? If they do not match, the evaluation criteria can be adjusted manually or automatically, e.g. using methods of reinforcement learning.

According to a preferred embodiment, the above pertinent information for training the machine learning system can be established based on human inspection of the component.

According to a further preferred aspect of the invention, the machine learning system can, in the classification step, analogously to the test methods described above, also classify into a "repairable" or "non-repairable" category, determine a probability of success of a repair and / or carry out an identification of flaws in the at least one recording of the component (10) and when entering the relevant information, a relevant category or correctly identified flaws are entered into the machine learning system (30) for training.

The object is further achieved by a computer program product according to claim 11, which has instructions readable by a processor of a computer which, when executed by the processor, cause the processor to execute the method according to one of the preceding aspects.

The object is further achieved by a computer-readable storage medium according to claim 12, on which the computer program product according to the preceding aspect is stored.

The object is also achieved by a system for testing a component, in particular a component of a turbomachine according to claim 13. Such a system has at least one image acquisition device for recording an X-ray or CT image of the component and a trained machine learning system which is designed to acquire/obtain image capture device image and metadata about the component and be trained to classify the component into a “servicable” or a “non-servicable” category based on that data. Furthermore, the system may be trained and configured to perform one or more of the aspects of a testing method or training method described above.

character list

The invention is explained in more detail below using an exemplary embodiment, with the individual features within the framework of the independent claims also being able to be essential to the invention in a different combination and no distinction being made in detail between the different claim categories.

• 1 eine schematische Darstellung einer Strömungsmaschine;
• 2 eine schematische Darstellung eines Ablaufs eines Verfahrens zum Prüfen eines Bauteils gemäß einem Ausführungsbeispiel der vorliegenden Erfindung;
• 3 eine schematische Darstellung eines Ablaufs eines Trainingsvorgangs für ein maschinelles Lernsystem gemäß einem Ausführungsbeispiel der Erfindung;
• 4 ein System gemäß einem Ausführungsbeispiel der Erfindung.

In detail shows

• 1 a schematic representation of a turbomachine;
• 2 a schematic representation of a sequence of a method for testing a component according to an embodiment of the present invention;
• 3 a schematic representation of a sequence of a training process for a machine learning system according to an embodiment of the invention;
• 4 a system according to an embodiment of the invention.

Preferred embodiment of the invention

1 shows a turbomachine 1, specifically a turbofan engine, in an axial section. The turbomachine 1 is functionally divided into a compressor 1a, a combustion chamber 1b and a turbine 1c. Both the compressor 1a and the turbine 1c are each constructed of multiple stages. Each of the stages is composed of a stator 5 and a rotor 6 . The reference numeral 7 designates the gas duct, ie the compressor gas duct in the case of the compressor 1a or the hot gas duct in the case of the turbine 1c. The sucked-in air is compressed in the compressor gas duct, it is then burned in the downstream combustion chamber 1b with added kerosene. The hot gas flows through the hot gas duct and drives the rotors 6 , which rotate about the axis of rotation 2 .

2 FIG. 12 illustrates, in a schematic representation, a method for testing a component 10, in particular a component 10 of a turbomachine, preferably a component 10 of an engine. according to an embodiment of the present invention. In a first step (S1), a component to be tested reaches a test station 100 (see 4 ). In a next step (S2), the component is placed in an image acquisition device 20 and at least one X-ray or CT image of the component is taken at a predetermined angle. The predetermined angle can, for example, be set in such a way that a sectional view of the component is produced, which highlights component areas that are particularly frequently damaged during operation.

In the subsequent method step (S3), the at least one recording captured by the image capture device 20 is transmitted to a machine learning system 30, which has a computer or a computing device on the at least one database and a trained self-learning algorithm are stored. The machine learning system 30 first determines whether the receptacle of the component or the component 10 has at least one defect, such as a crack or at least one pore, or not. In addition, the machine learning system 30 can output or generate an image on which the detected or identified defects, in particular cracks and/or pores, in the recording of the component 10 are marked. These markings can include, for example, colored elements that correspond to the shape of the defect, the crack or the pore, inscriptions or the like and serve as a guide for the tester/assessor in order to mark defects on the actual component. If the machine learning system 30 has not recognized any defect or no crack and/or no pore in the overall image, the component 10 can be identified as being free of defects.

In the exemplary embodiment described, the algorithm is trained to classify the component 10 either in a “servicable” or in a “non-servicable” category. "Serviceable" means that the component could be put into operation in accordance with the regulations. This means that the component was either found to be free of defects or the flaws determined by the machine learning system 30 are so minor that no effects on the component 10 during operation are to be expected, even with a high safety factor involved. In order to carry out the classification, the trained, self-learning algorithm accesses the metadata stored in the database for component 10, which includes the running time, the remaining life cycles (before repair), the operator (e.g. airline), nominal or measured wall thicknesses, a repair history or include a history of the environment in which the device was operated.

Step (S4) is only carried out if the component 10 was classified as “non-servicable” in step (S3). In this case, the machine learning system makes a further classification of the component: either into a "repairable" or into a "non-repairable" category. If the component is classified in the "Repairable" category, the machine learning system creates a forecast of the probability of success of the repair in a step (S5). For this prognosis, the machine learning system 30 can access a database in which both previous test methods/results on components of the same type (recorded images, metadata and test result) and the final results of the repairs of these components are stored. In other words, image and metadata created in the past for components of the same type, the forecasts made using this data and the ultimate repair success or failure are linked in the database in order to serve as a database for a forecast of a currently tested part. For example, if a crack of a certain length was found in a component currently being examined, the database could be used to check how high the probability of success in repairing the same type of component with a crack of comparable length (e.g. +/- 10%) was in the past .

Once step S5 has been completed, the component 10 is forwarded to the next station for the first repair step with the prognosis made about the probability of recovery and the determined error positions, types and progress variables.

The length of the defect or defects, in particular the crack or cracks or the pore or pores, the width of the defect or defects, in particular the crack or of the cracks, the size or diameter of the defect or pore and/or the number of defects, in particular cracks and/or pores, are taken as a basis for the classification property.

• Als Ausgangsbasis für das Training des maschinellen Lernsystems 30 ist denkbar, dass Bilder, die vom Menschen zutreffend als Fehlstellen, insbesondere Risse und/oder Poren, „aufweisend“ (auch als „eine Anzeige haben“ bezeichnet) oder „nicht aufweisend“ (auch als „keine Anzeige haben“ bezeichnet) markiert wurden, in das maschinelle Lernsystem 30 als Truth/Wahrheit eingegeben werden. Vorstellbar ist auch, dass zusätzlich das Bild, bei dem ein Mensch die Stellen bzw. Positionen der Fehlstellen, insbesondere Risse und/oder Poren, zutreffend markiert hat, als Truth/Wahrheit in das maschinelle Lernsystem 30 eingegeben wird. Diese Bilder werden im Kontext mit den jeweiligen bereits erwähnten Metadaten (Laufzeit, LLP etc.) in das maschinelle Lernsystem 30 eingegeben. Zusammengefasst kann ein vom Menschen vorbewerteter Datensatz als Ausgangsbasis für das Anlernen des maschinellen Lernsystems 30 dienen.

3 As already mentioned, the machine learning system 30 has a trained self-learning algorithm. This can in particular be a support vector machine (SVM), a neural network, a deep neural network (DNN), a convolutional neural network (CNN) or the like. This will, as in 3 shown, trained to test components as follows:

• As a starting point for the training of the machine learning system 30, it is conceivable that images that humans correctly identify as defects, in particular cracks and/or pores, "have" (also referred to as "having a display") or "not exhibiting" (also referred to as " labeled "have no display") are input into the machine learning system 30 as truth. It is also conceivable that the image in which a person has correctly marked the locations or positions of the defects, in particular cracks and/or pores, is also entered into the machine learning system 30 as truth. These images are entered into the machine learning system 30 in context with the respective metadata already mentioned (runtime, LLP etc.). In summary, a human pre-scored data set can be used as a Starting point for teaching the machine learning system 30 are used.

After that, the machine learning system 30 can be further trained in situ, so to speak, and its error detection can be sharpened. An image or an image recording of the component 10 that was created by the image acquisition device 20 is input into the machine learning system 30 for this purpose. The component 10 was X-rayed during the image recording, so that a gray scale representation is produced in which flaws or irregularities lying below the surface of the component 10 are also contained in the signal. The machine learning system 30 outputs the information as to whether the component 10 has defects, in particular cracks or pores, or not, based on the image recorded by the image acquisition device 20, the metadata stored in the database for the component 10 and the learning history, and outputs an image on which the defects, in particular cracks and/or pores, which it has identified are marked. In addition, the machine learning system 30 classifies according to the in 2 described procedure, the component 10 based on the above data as servicable / non-servicable and repairable / non-repairable.

A trainer, e.g. a human assessor, checks the recording of the component (and possibly also the physical component itself) at the same time or subsequently and corrects or correctly states whether the respective component contains 10 defects, in particular cracks and/or pores or not. This is fed into the machine learning system 30 so that the machine learning system 30 can learn. It is also possible, for example, for the human appraiser/trainer to manually mark a defect, such as a crack, in order to train the algorithm. The trainer also gives his classification of the component 10 in the categories servicable/non-servicable and repairable/non-repairable as a reference.

The machine learning system 30 independently recognizes the differences between the image it generates, on which the defects recognized by the machine learning system 30, in particular cracks and/or pores, are marked, and the image on which the defects recognized correctly by the human, in particular Cracks and/or pores are marked. The machine learning system also independently recognizes deviations in its classification of the component for human classification. If the machine learning system 30 has not recognized a defect or incorrectly identified a non-defective area as a defect, it learns on the basis of the human finding and can reweight its evaluation criteria.

If the machine classification and error detection agree with the trainer's classification and error detection, the machine learning system 30 is further trained on the next component. If there is a discrepancy, the trainer compares the results and can manually adjust the criteria. Alternatively, the machine learning system can also automatically adjust its evaluation criteria based on the human evaluation. The recording is then reevaluated by the machine learning system 30 using the adjusted criteria. This is repeated until the machine score matches the human score.

This training of the machine learning system 30 can be repeated with a large number of images, if possible of different components 10 of the same type.

The machine learning system 30 can be trained by different people. This means that different people identify flaws, in particular cracks and/or pores, in the overall image and this information is entered into the machine learning system 30 .

If the machine learning system 30 has been trained to such an extent that it carries out the search for defects and the classification of the component as well as a person (trained or trained for this) the training can be ended if necessary and the person only has to operate the system 10 be monitored or operated with the machine learning system 40 .

For example, the component 10 may be a component of an engine. For example, the component 10 may be a rotor or blade of a high pressure compressor (HPC) or a rotor or blade of a high pressure turbine (HPT) of an engine. The engine can in particular be an engine of an aircraft.

Reference List

1

flow machine

1a

compressor

1b

combustion chamber

1c

turbine

2

axis of rotation

5

rotor

6

stator

7

gas channel

10

component

20

image capture device

30

machine learning system

100

test station

Claims (13)

Method for testing a component (10), in particular a turbomachine (1), with the steps: - Creating (S2) at least one X-ray or CT image of the component (10) by means of an image acquisition device (20); - Providing (S21) metadata about the component (10), which include in particular a component type, a term of the component (10), a number of remaining life cycles and/or a repair history; - Classifying the component (10) based on the image created by the image acquisition device (20) and the metadata provided in a “servicable” or a “non-servicable” category using a trained machine learning system (30). procedure after claim 1 , in which the machine learning system (30) assigns the parts classified as “non-serviceable” to either a “repairable” or a “non-repairable” category. procedure after claim 2 , in which the machine learning system (30) assigns a probability of success of the repair to the components classified as “repairable”. Method according to one of the preceding claims, in which the machine learning system (30) comprises a neural network, a deep neural network (DNN), a convolutional neural network (CNN) and/or a support vector machine. Method according to one of the preceding claims, in which the machine learning system (30) is designed to identify and/or localize defects, in particular cracks or pores, in the at least one recording and to identify and/or localize the identified/localized defects, in particular a type , number and/or size of the identified defects must be taken into account in the classification of the component. Method according to one of the preceding claims, in which the machine learning system (30) is designed to place a filter over the recording obtained from the image acquisition device (20) as a step of the classification. Method according to one of the preceding claims, in which the machine learning system (30) is designed to autonomously control the image acquisition device after the evaluation of the at least one recording in order to create at least one further recording of the component with a varied recording parameter, in particular a varied recording angle , if a classification criterion cannot be met based on the at least one initial recording. Method for training a machine learning system (30) for testing a component (20), in particular a component (20) of a turbomachine, the method comprising the following steps: Providing a machine learning system (30), in particular comprising a neural network; Entering a recording of the component (10) into the machine learning system (30), Entering metadata about the component (10) into the machine learning system (30), which includes in particular a component type, a term of the component (10), a number of remaining life cycles and/or a repair history Classify the component based on the entered data into a "servicable" or a "non-servicable" category; outputting the determined category; and inputting pertinent information about the category of the component (10) into the machine learning system (30) to train the machine learning system (30). procedure after claim 8 , wherein the correct information is created on the basis of a human examination of the component (10). procedure after claim 8 or 9 , wherein the machine learning system (30) in the step of classification additionally classifies into a “repairable” or “non-repairable” category, determines a probability of success of a repair and/or identifies defects in the at least one recording of the component ( 10) and when entering the correct information, a correct category or correctly identified defects for training are entered into the machine learning system (30). A computer program product comprising instructions readable by a processor of a computer which, when executed by the processor, cause the processor to carry out the method according to any one of the preceding claims. Computer-readable medium on which the computer program product claim 11 is saved. System (100) for testing a component (10), in particular a component (10) of a turbomachine, the system comprising the following: an image acquisition device (20) for recording an X-ray or CT image of the component (10); and a trained machine learning system (40) which is designed to receive the recording of the image acquisition device (20) and metadata about the component (10) and which is trained to convert the component (10) into a "servicable" or to be classified in a non-servicable category.

Images